/**
* @brief Set the inputs of the network from a patch within an image.
* It is assumed that the image is mono (1 byte per pixel)
* @param net Backprop neural net object
* @param img Array storing the image
* @param image_width Width of the image in pixels
* @param image_height Height of the image in pixels
* @param tx Top left x coordinate of the patch within the image
* @param ty Top left y coordinate of the patch within the image
*/
void bp_inputs_from_image_patch(bp * net,
unsigned char * img,
int image_width, int image_height,
int tx, int ty)
{
int px,py,i=0,idx;
/* The patch size is calculated from the number of inputs
of the neural net. It's assumed to be square. */
int patch_size = (int)sqrt(net->NoOfInputs);
/* make sure that the patch fits within the number of inputs */
assert(patch_size*patch_size <= net->NoOfInputs);
/* set the inputs from the patch */
for (py = ty; py < ty+patch_size; py++) {
if (py >= image_height) break;
for (px = tx; px < tx+patch_size; px++, i++) {
if (px >= image_width) break;
/* array index within the image */
idx = (py*image_width) + px;
/* set the input value within the range 0.25 to 0.75 */
bp_set_input(net, i, 0.25f + (img[idx]*0.5f/255.0f));
}
}
}
static void test_backprop_update()
{
bp net;
int no_of_inputs=10;
int no_of_hiddens=4;
int hidden_layers=2;
int no_of_outputs=5;
int i;
unsigned int random_seed = 123;
printf("test_backprop_update...");
bp_init(&net,
no_of_inputs, no_of_hiddens,
hidden_layers,
no_of_outputs, &random_seed);
assert((&net)->inputs!=0);
assert((&net)->hiddens!=0);
assert((&net)->outputs!=0);
/* set some inputs */
for (i = 0; i < no_of_inputs; i++) {
bp_set_input(&net, i, i/(float)no_of_inputs);
}
for (i = 0; i < 100; i++) {
bp_update(&net,0);
}
bp_free(&net);
printf("Ok\n");
}
static void test_backprop()
{
bp net;
int no_of_inputs=10;
int no_of_hiddens=4;
int hidden_layers=2;
int no_of_outputs=5;
int i,l;
unsigned int random_seed = 123;
printf("test_backprop...");
bp_init(&net,
no_of_inputs, no_of_hiddens,
hidden_layers,
no_of_outputs, &random_seed);
assert((&net)->inputs!=0);
assert((&net)->hiddens!=0);
assert((&net)->outputs!=0);
/* set some inputs */
for (i = 0; i < no_of_inputs; i++) {
bp_set_input(&net, i, i/(float)no_of_inputs);
(&net)->inputs[i]->BPerror = 999;
}
for (l = 0; l < hidden_layers; l++) {
for (i = 0; i < bp_hiddens_in_layer(&net,l); i++) {
(&net)->hiddens[l][i]->BPerror = 999;
}
}
/* set some target outputs */
for (i = 0; i < no_of_outputs; i++) {
(&net)->outputs[i]->BPerror = 999;
bp_set_output(&net, i, i/(float)no_of_inputs);
}
/* feed forward */
bp_feed_forward(&net);
bp_backprop(&net,0);
/* check for non-zero backprop error values */
for (i = 0; i < no_of_inputs; i++) {
assert((&net)->inputs[i]->BPerror != 999);
}
for (l = 0; l < hidden_layers; l++) {
for (i = 0; i < bp_hiddens_in_layer(&net,l); i++) {
assert((&net)->hiddens[l][i]->BPerror != 999);
}
}
for (i = 0; i < no_of_outputs; i++) {
assert((&net)->outputs[i]->BPerror != 999);
}
bp_free(&net);
printf("Ok\n");
}
/**
* @brief Convolve an array of floats to the input layer of a neural net
* @param samples_across The number of units across in the layer of neurons
* (sampling grid resolution)
* @param samples_down The number of units down in the layer of neurons
* (sampling grid resolution)
* @param patch_radius The radius of the patch within the float array
* @param floats_width Width of the image
* @param floats_height Height of the image
* @param floats_depth Depth of the image (mono=1, RGB=3)
* @param layer0 Array of floats
* @param net Neural net to set the inputs for
* @param feature_autocoder An autocoder containing learned features
* @param use_dropouts Non-zero if dropouts are to be used
* @returns zero on success
*/
int features_conv_floats_to_neurons(int samples_across,
int samples_down,
int patch_radius,
int floats_width,
int floats_height,
int floats_depth,
float layer0[],
bp * net,
ac * feature_autocoder,
unsigned char use_dropouts)
{
int no_of_learned_features = feature_autocoder->NoOfHiddens;
if (samples_across * samples_down * no_of_learned_features !=
net->NoOfInputs) {
/* across*down doesn't equal the second layer units */
return -1;
}
if (feature_autocoder->NoOfInputs !=
patch_radius*patch_radius*4*floats_depth) {
/* the patch size doesn't match the feature
learner inputs */
return -2;
}
for (int fy = 0; fy < samples_down; fy++) {
for (int fx = 0; fx < samples_across; fx++) {
int tx=0, ty=0, bx=0, by=0;
if (features_patch_coords(fx, fy,
samples_across, samples_down,
patch_radius,
floats_width, floats_height,
&tx, &ty, &bx, &by) != 0) {
continue;
}
if (scan_floats_patch(layer0,
floats_width, floats_depth,
tx, ty, bx, by,
feature_autocoder) != 0) {
return -4;
}
int index_net_inputs =
(fy * samples_across + fx) *
no_of_learned_features;
autocoder_encode(feature_autocoder, feature_autocoder->hiddens,
use_dropouts);
for (int f = 0; f < no_of_learned_features; f++) {
bp_set_input(net, index_net_inputs+f,
autocoder_get_hidden(feature_autocoder, f));
}
}
}
return 0;
}
static void test_backprop_autocoder()
{
bp autocoder;
int itt,i,j;
int no_of_inputs=10;
int no_of_hiddens=4;
int no_of_outputs=10;
unsigned int random_seed = 123;
float tot;
printf("test_backprop_autocoder...");
/* create the autocoder */
bp_init(&autocoder,
no_of_inputs,
no_of_hiddens,1,
no_of_outputs,
&random_seed);
autocoder.learningRate = 0.5f;
/* run the autocoder for some itterations */
for (itt = 0; itt < 100; itt++) {
/* set the inputs */
for (i = 0; i < no_of_inputs; i++) {
bp_set_input(&autocoder,i,0.25f + (i*0.5f/(float)no_of_inputs));
bp_set_output(&autocoder,i,0.75f - (i*0.5f/(float)no_of_inputs));
}
/* update */
bp_update(&autocoder,0);
}
for (i = 0; i < no_of_hiddens; i++) {
/* check that some errors have been back-propogated */
assert((&autocoder)->hiddens[0][i]->BPerror != 0);
/* check that weights have changed */
tot = 0;
for (j = 0; j < no_of_inputs; j++) {
assert((&autocoder)->hiddens[0][i]->lastWeightChange[j]!=0);
tot += fabs((&autocoder)->hiddens[0][i]->lastWeightChange[j]);
}
/* total weight change */
assert(tot > 0.00001f);
}
bp_free(&autocoder);
printf("Ok\n");
}
/**
* @brief Pre-trains a hidden layer using an autocoder
* @param net Backprop neural net object
* @param autocoder Autocoder neural net object
* @param hidden_layer The layer within the neural net to be autocoded
*/
void bp_pretrain(bp * net,
bp * autocoder,
int hidden_layer)
{
int i;
float hidden_value;
/* feed forward to the given hidden layer */
bp_feed_forward_layers(net, hidden_layer);
if (hidden_layer > 0) {
/* check that the number of inputs is valid */
assert(net->NoOfHiddens == autocoder->NoOfInputs);
/* copy the hidden unit values to the inputs
of the autocoder */
for (i = 0; i < net->NoOfHiddens; i++) {
hidden_value = bp_get_hidden(net, hidden_layer-1, i);
bp_set_input(autocoder,i, hidden_value);
}
}
else {
/* check that the number of inputs is valid */
assert(autocoder->NoOfInputs == net->NoOfInputs);
/* copy the input unit values to the inputs
of the autocoder */
for (i = 0; i < net->NoOfInputs; i++) {
bp_set_input(autocoder, i,
bp_get_input(net, i));
}
}
/* run the autocoder */
bp_update_autocoder(autocoder);
}
/**
* @brief Set the inputs of the network from an image.
* It is assumed that the image is mono (1 byte per pixel)
* @param net Backprop neural net object
* @param img Array storing the image
* @param image_width Width of the image in pixels
* @param image_height Height of the image in pixels
*/
void bp_inputs_from_image(bp * net,
unsigned char * img,
int image_width, int image_height)
{
int i=0;
/* check that the number of inputs is the same as the
number of pixels */
assert(net->NoOfInputs == image_width*image_height);
/* set the inputs */
for (i = 0; i < image_width*image_height; i++) {
/* set the input value within the range 0.25 to 0.75 */
bp_set_input(net, i, 0.25f + (img[i]*0.5f/255.0f));
}
}
static void test_backprop_deep()
{
bp net;
bp autocoder;
int l,itt,i;
int no_of_inputs=10;
int no_of_hiddens=4;
int hidden_layers=3;
int no_of_outputs=2;
unsigned int random_seed = 123;
printf("test_backprop_deep...");
bp_init(&net,
no_of_inputs, no_of_hiddens,
hidden_layers,
no_of_outputs, &random_seed);
assert((&net)->inputs!=0);
assert((&net)->hiddens!=0);
assert((&net)->outputs!=0);
for (l = 0; l < hidden_layers; l++) {
/* create an autocoder for this layer */
bp_create_autocoder(&net,l,&autocoder);
/* do some training */
for (itt = 0; itt < 100; itt++) {
/* set the inputs */
for (i = 0; i < no_of_inputs; i++) {
bp_set_input(&net,i,i/(float)no_of_inputs);
}
/* update */
bp_pretrain(&net,&autocoder,l);
}
/* move the autocoder hidden weights into the main network */
bp_update_from_autocoder(&net,&autocoder,l);
/* delete the autocoder */
bp_free(&autocoder);
}
bp_free(&net);
printf("Ok\n");
}
static void test_backprop_feed_forward()
{
bp net;
int no_of_inputs=10;
int no_of_hiddens=4;
int hidden_layers=2;
int no_of_outputs=5;
int i;
unsigned int random_seed = 123;
printf("test_backprop_feed_forward...");
bp_init(&net,
no_of_inputs, no_of_hiddens,
hidden_layers,
no_of_outputs, &random_seed);
assert((&net)->inputs!=0);
assert((&net)->hiddens!=0);
assert((&net)->outputs!=0);
/* set some inputs */
for (i = 0; i < no_of_inputs; i++) {
bp_set_input(&net, i, i/(float)no_of_inputs);
}
/* clear the outputs */
for (i = 0; i < no_of_outputs; i++) {
(&net)->outputs[i]->value = 999;
}
/* feed forward */
bp_feed_forward(&net);
/* check for non-zero outputs */
for (i = 0; i < no_of_outputs; i++) {
assert((&net)->outputs[i]->value != 999);
}
bp_free(&net);
printf("Ok\n");
}
static void test_backprop_training()
{
bp * net;
int no_of_inputs=2;
int no_of_hiddens=2;
int hidden_layers=1;
int no_of_outputs=1;
int itt,example;
unsigned int random_seed = 123;
float state_TRUE = 0.8f;
float state_FALSE = 0.2f;
printf("test_backprop_training...");
net = (bp*)malloc(sizeof(bp));
bp_init(net,
no_of_inputs, no_of_hiddens,
hidden_layers,
no_of_outputs, &random_seed);
assert(net->inputs!=0);
assert(net->hiddens!=0);
assert(net->outputs!=0);
/* training */
example=0;
for (itt = 0; itt < 500000; itt++, example++) {
if (example>=4) example=0;
/* select an example from the XOR truth table */
switch(example) {
case 0: {
bp_set_input(net, 0, state_FALSE);
bp_set_input(net, 1, state_FALSE);
bp_set_output(net, 0, state_FALSE);
break;
}
case 1: {
bp_set_input(net, 0, state_TRUE);
bp_set_input(net, 1, state_FALSE);
bp_set_output(net, 0, state_TRUE);
break;
}
case 2: {
bp_set_input(net, 0, state_FALSE);
bp_set_input(net, 1, state_TRUE);
bp_set_output(net, 0, state_TRUE);
break;
}
case 3: {
bp_set_input(net, 0, state_TRUE);
bp_set_input(net, 1, state_TRUE);
bp_set_output(net, 0, state_FALSE);
break;
}
}
/* train on the example */
bp_update(net,0);
}
bp_set_input(net, 0, state_FALSE);
bp_set_input(net, 1, state_FALSE);
bp_feed_forward(net);
if (bp_get_output(net, 0) >= 0.5f) {
printf("\n%.5f\n",bp_get_output(net, 0));
}
assert(bp_get_output(net, 0) < 0.5f);
bp_set_input(net, 0, state_FALSE);
bp_set_input(net, 1, state_TRUE);
bp_feed_forward(net);
if (bp_get_output(net, 0) <= 0.5f) {
printf("\n%.5f\n",bp_get_output(net, 0));
}
assert(bp_get_output(net, 0) > 0.5f);
bp_set_input(net, 0, state_TRUE);
bp_set_input(net, 1, state_FALSE);
bp_feed_forward(net);
if (bp_get_output(net, 0) <= 0.5f) {
printf("\n%.5f\n",bp_get_output(net, 0));
}
assert(bp_get_output(net, 0) > 0.5f);
bp_set_input(net, 0, state_FALSE);
bp_set_input(net, 1, state_FALSE);
bp_feed_forward(net);
if (bp_get_output(net, 0) >= 0.5f) {
printf("\n%.5f\n",bp_get_output(net, 0));
}
assert(bp_get_output(net, 0) < 0.5f);
bp_free(net);
free(net);
printf("Ok\n");
}
